An event causal relationship identification method fusing argument and structure information

Abstract

The application discloses an event causal relationship identification method fusing argument and structure information, and has the characteristics that a multi-task learning framework is adopted, and basic argument and structure information perception causal question answers are used to enhance event causal relationship identification; document-level event causal relationship identification is converted into a multiple-choice question generation type question answering, and argument and structure information are fused, and a large language model is used to generate the causal relationship of a questioned event, and the method specifically comprises the following steps: data preparation, argument generation, structure information and model training and the like. Compared with the prior art, the application has the advantages of high identification accuracy and efficiency, brings a new frontier and breakthrough to the field of document-level event causal relationship identification, is simple in method, good in use effect, suitable for limited computing cost and acceptable time, and has good application prospect and commercial value.

Description

Technical Field

[0001] This invention relates to the field of information extraction and event causality recognition technology, specifically a method for identifying event causal relationships based on a multi-task learning framework that integrates arguments and structural information. Background Technology

[0002] Document-level event causality identification (DECI) aims to identify the causal relationship between two events in a document, and is an important yet challenging task in information extraction. Research on causal relationship extraction helps us solve many practical problems, such as the automatic processing of massive amounts of information, thereby supporting a wide range of downstream applications, including future event prediction, knowledge graph construction, and global crisis monitoring. Figure 1 As shown, the event "eruption" caused the "tsunami," so the two events can be considered to have a causal relationship. The mention of "waves" and the mention of "tsunami" in both events refer to the same event, therefore they can be considered to have a coreference relationship.

[0003] Recent advances in large language models (LLMs) have sparked considerable interest in applying these models to various language tasks, such as model editing, model extraction, and efficient parameter fine-tuning. However, in the task of causal event extraction, recent research has tended to use pre-trained language models to generate event causal relationships. However, due to the multiple events in a document, these methods are prone to sequential generation errors, have limited performance, and are costly. In an era of rapid development in large language models, the potential of these traditional methods is severely limited, and achieving significant breakthrough performance, even with added components, appears to be a formidable challenge.

[0004] The paper "Event causality identification via generation of important context words" proposes GenECI, which uses a generative model (T5) to infer causal relationships between events by capturing key context words. However, GenECI uses true or false questions to determine the causal relationship between two events, leading to high time costs and being only applicable to sentence-level extraction. The paper "SEAG: Structure-Aware EventCausality Generation" proposes SEAG, which combines event detection and event causal relationship identification, demonstrating the effectiveness of linearizing the event causal graph structure through linearization and contrastive learning methods. However, SEAG is not suitable for causal extraction of labeled events, and simply labeling event words or restating events is unlikely to effectively improve its performance in causal extraction of known events. Furthermore, existing generative methods generate results only through supervised templates, and the language model may not understand why such extraction is correct. Therefore, in practice, how to effectively and evidence-basedally extract causal relationships within constraints of time and computational cost remains a challenge that needs to be addressed.

[0005] Existing technologies for event causality recognition rely on pre-trained encoder models and graph neural network variants, using pre-trained language models to generate event causality relationships, which is time-consuming. Using pre-trained language models to generate event causality relationships is prone to generation errors due to the multiple events in a document. Furthermore, it neglects potential structures such as event coreference and related causal chains. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for identifying event causal relationships that integrates argumentation and structural information. It employs a multi-task learning framework and constructs a linearized event causal relationship structure graph. Using a decoder-only generative large language model, causal relationship extraction is transformed into a task solvable through natural language generation. Event causal relationship identification is enhanced by answering causal questions perceived through basic arguments and structural information. The causal identification between event pairs is converted into a multiple-choice question. For each event, referred to as the "observed event," the question asks which of the listed options represents its causal relationship. Reliable arguments provided by the intelligent model are combined with the constructed linearized event causal relationship graph to extract important contextual key information and enhance the indirect supervision of important event relationship references, thereby improving the accuracy of selection. This method improves the accuracy and efficiency of identification by converting the DECI task into a multiple-choice question and utilizing a large language model to generate event causal relationships. Furthermore, this method provides arguments to explain event causal relationships, enhancing the interpretability of the model. It introduces an event structure graph to model multi-hop potential relationships in causal reasoning, thus more comprehensively capturing the causal connections between events, bringing a new frontier and breakthrough to the field of document-level event causal relationship identification. The method is simple, effective, and adaptable to limited computing costs and acceptable timeframes, making it a promising application with significant commercial value.

[0007] The specific technical solution to achieve the purpose of this invention is: a method for identifying event causal relationships by integrating arguments and structural information, characterized by: cropping the text to be asked; constructing options for the questions on the training set and the test set; using a multi-task training method; integrating the arguments generated by the intelligent model and the constructed structural information into the multiple-choice QA question; performing multi-task training; and fine-tuning the model, specifically including the following steps:

[0008] (I) Data Preparation

[0009] 1-1: Convert the causal identification between event pairs into a multiple-choice question. For each event, which is called the "observed event", ask which of the listed options is the causal relationship.

[0010] 1-2: Trim and prune the text to be addressed, retaining only the beginning and end of the observed events and candidate events to reduce noise in the text and eliminate the influence of irrelevant events.

[0011] 1-3: Construct question options for the training and test sets. For the training set, all events with causal relationships are included in the options. If coreference events exist, the one closest to the observed event is selected, and at least three unrelated events are added as distractors. For the test set, each event is treated as an observed event and iterated sequentially, with each event set mentioned as an option. The training and test sets are derived from an open-source dataset and obtained by splitting the training and test sets in an 8:2 ratio.

[0012] 1-4: The training set constructed above is used to train the multiple-choice QA model. The multiple-choice QA model is based on the open-source Baichuan2-7B-Chat generation model, and generating answers based on questions is our main task. The test set is used to evaluate the performance of the multiple-choice QA model.

[0013] (II) Arguments for Integrating Intelligent Model Generation

[0014] Arguments from smarter models are artificially obtained and injected into the training model as one of the auxiliary tasks. When the model is trained, it serves as a supervisory tool for predicting answers. The reliable arguments provided by smarter models are integrated to extract important contextual key information.

[0015] (III) Integration of Structural Information

[0016] The events are sorted according to their location, and the event coreference and causal relationship are combined in a more natural language way. The linearization method of causal graph structure is used to obtain the event causal graph structure described in natural language. The linearized event causal relationship graph constructed by fusion is used as another auxiliary task to enhance the information mentioned in the important event relationship as indirect supervision.

[0017] (iv) Model training and optimization

[0018] The arguments generated by the fusion intelligent model and the structural information constructed are incorporated into the multiple-choice QA model. Multi-task training is performed to fine-tune the model. The same generative model learns three tasks simultaneously and is trained to minimize the loss.

[0019] Compared with existing technologies, this invention transforms the DECI task into a multiple-choice question, utilizes a large-scale language model to generate causal relationships between events, and improves the accuracy and efficiency of recognition. Furthermore, this method provides arguments to explain the causal relationships between events, enhancing the interpretability of the model. It introduces an event structure graph to model multi-hop potential relationships in causal reasoning, thereby more comprehensively capturing the causal connections between events. This brings a new frontier and breakthrough to the field of document-level event causal relationship recognition. The method is simple, effective, and adaptable to limited computational costs and acceptable timeframes, demonstrating promising application prospects and commercial value. Attached Figure Description

[0020] Figure 1 This is a diagram of an example. Detailed Implementation

[0021] The following example of a reported tsunami event further illustrates the invention:

[0022] Example 1

[0023] See Figure 1 A naval report states that an underwater volcanic eruption triggered a tsunami, striking an archipelago with 10-meter-high waves. Police confirmed two deaths in the tsunami. However, due to poor communication, it is currently unclear whether there were more casualties. This invention employs a multi-task learning framework to address this event, integrating arguments generated by an intelligent model and constructed structural information into a multiple-choice QA question, perceiving causality through basic arguments and structural information.

[0024] Answering questions to enhance the identification of causal relationships in events involves the following steps:

[0025] (I) Data Preparation

[0026] 1-1: Convert the causal identification between event pairs into a multiple-choice question. For each event as an unobserved event, ask which of the listed options represents its causal relationship.

[0027] 1-2: Text construction and trimming of the title text, including only the observed events and candidate events, with head and tail trimming to reduce redundant text noise and avoid the influence of irrelevant events, even if this may result in partial semantic loss of text with abnormal narrative order.

[0028] 1-3: Constructing question options for the training and test sets, i.e., building a set of candidate events that match the questions as multiple-choice options. The data for the training and test sets comes from two widely used open-source datasets: EventStoryLine and Causal-TimeBank. EventStoryLine contains 22 topics, 258 documents, 4316 sentences, and 5334 event mentions, totaling 54326 event pairs, of which 5655 event pairs have causal relationships. Causal-TimeBank contains 184 documents and 6813 events, of which 318 events have causal relationships with the annotations. 80% of this is used as the training set, and the remaining 20% ​​is used as the test set.

[0029] The training set adds all causally related events to the options. If coreference exists, only the one closest to the observed event is selected, followed by at least three irrelevant events as distractors. This approach reduces text length and allows for understanding as many relevant causal event pairs as possible. Furthermore, in contrastive learning, distractors can be viewed as a form of negative sample, helping the model distinguish the relationships between different candidate events and the observed event. In summary, these options can be described as (1) relevant event mentions, (2) irrelevant distractors, and (3) "none of the above".

[0030] The test set iterates over each event as an observed event, and each event set is mentioned as an option, which can be described as (1) candidate event, (2) "none of the above".

[0031] The training set constructed above is used to train a multiple-choice QA model, where generating answers based on questions is our main task. The input is the question, and the output is option labels, such as ABCD. The test set constructed above is used to evaluate the performance of the multiple-choice QA model in causal relationship recognition.

[0032] (II) Generation of Arguments

[0033] Arguments generated by the fusion intelligent model are injected into the aforementioned constructed multiple-choice QA model as an auxiliary task. The input is the question, and the output is the argument. During the model's pre-training phase, multiple-choice questions (MCQs) are commonly used as both the dataset and the task format. Many pre-trained MCQ datasets include parsing, which provides the reasoning behind choosing the answer. Therefore, arguments can be artificially injected as a supervisory tool for predicting answers during model training.

[0034] (III) Construction of Structural Information

[0035] Structural information is injected into the constructed multiple-choice QA model as another auxiliary task, with the question as input and structural information as output. Events are ordered according to their position, and event coreference and causal relationships are combined in a natural language manner. A causal graph structure linearization method is then used to obtain an event causal graph structure described in natural language.

[0036] The specific steps of step (iii) are as follows:

[0037] A series of previous works have demonstrated that graph neural networks are effective in predicting causal relationships because they aggregate features of neighboring nodes (where nodes are events), which can enhance the model’s judgment of causal transitivity and event coreference resolution to some extent.

[0038] See Figure 1 d, ECG display Figure 1 Within the dashed box in (d), as part of the enhanced representation, is a graph G = (E, V). Inspired by SEAG, it is ordered according to the location of events. To combine event coreference and causality in a more natural way, G is linearized as shown in equation (a) below:

[0039] g = [e1, r coref ,e2,…,e1,r cause [e3, ...] (a)

[0040] Where g represents the linearized structural information, e i Let r represent the i-th event. coref The conjunction r indicates co-reference. cause Words indicating cause and effect.

[0041] First, all the identical events are mentioned, then all causal relationships are presented. In the experiment, "is" is used as r. coref "It is because of..." as r cause Each relation is repeated in reverse, e.g., e1 is the cause of e3, e3 is the effect of e1, ...}. It's important to note that the linearized ECG is only related to the segmented text, and not to the observed events or options.

[0042] (iv) Model training and optimization

[0043] The arguments generated by the fusion intelligent model and the structural information constructed are injected into the multiple-choice QA model for multi-task training. The trained multiple-choice QA model is then fine-tuned. The same generated model learns three tasks simultaneously and is trained to minimize the loss. The final multiple-choice QA model is used to identify the causal relationship of events.

[0044] The specific steps of step (iv) are as follows:

[0045] The generative model f (i.e., a transformer-based multiple-choice QA model) learns three tasks simultaneously and is trained using equations (b) to (e) below to minimize the following label prediction loss:

[0046]

[0047]

[0048]

[0049]

[0050] Where, α tThe weights represent the loss between tasks, where l is the cross-entropy loss between the predicted token and the target token, (p t q) represents the concatenation of custom prompts and questions as input to the model, a is the label of the correct answer, r is the argument generated by the intelligent model, g is the structural information of the construction, t is the specific task, T is the task set, Q→A, Q→R, Q→G correspond to the loss L label L rationale L ecg The prompt p predict p explain p extract And output a, r, g respectively.

[0051] See Figure 1 b-c, the present invention transforms document-level event causality identification (DECI) into a multiple-choice generative question answering method, and improves it by integrating evidence and structural information. Examples of prompts and outputs are shown in Figure 2(b)(c). DECI aims to identify the causal relationship between two events in a document.

[0052] This invention proposes a multi-task learning framework to enhance event causality identification through causal question answering that is aware of basic arguments and structural information. Specifically, the DECI task is transformed into multiple-choice question answering, and a large-scale language model is used to generate causal relationships for the challenged events. Furthermore, arguments explaining why these events are causally related are proposed, and an event structure graph is constructed that models the multi-hop potential relationships in the causal reasoning of the current event. Compared with state-of-the-art methods, this invention has significant advantages.

[0053] The above specific implementations are merely further illustrations of the present invention and are not intended to limit the scope of the patent. All equivalent implementations of the present invention should be included within the scope of the claims of the patent.

Claims

1. A method for identifying event causal relationships by integrating evidence and structural information, characterized in that, A multi-task learning framework is adopted to integrate arguments generated by intelligent models and structural information into multiple-choice QA questions. This enhances the identification of causal relationships by using answers to causal questions perceived through basic arguments and structural information. Specifically, the following steps are included: (a) Data preparation 1-1: Convert the causal identification between event pairs into a multiple-choice question. For each event as an unobserved event, ask which of the listed options represents its causal relationship. 1-2: Trim the problem text, retaining only the beginning and end of the observed events and candidate events to reduce noise in the text and eliminate the influence of irrelevant events; 1-3: Construct question options for the training and test sets, i.e., construct a set of candidate events that match the question as multiple-choice options. The training set includes all events with causal relationships as options. If there are co-referenced events, the one closest to the observed event is selected, and at least three unrelated events are added as distractors. The test set iterates through each event as an observed event, and each event set is mentioned as an option. The training and test sets are derived from open-source datasets and obtained by splitting the training and test sets in an 8:2 ratio. 1-4: The training set constructed above is used to train the multiple-choice QA model, and the test set is used to evaluate the performance of the multiple-choice QA model in causal relationship recognition. The multiple-choice QA model is based on the open-source Baichuan2-7B-Chat generation model, whose input is the question and output is the option label. (ii) Generation of arguments Arguments generated by the integrated intelligent model are injected into the multiple-choice QA model constructed above, serving as a supervisory tool for predicting answers during model training; (III) Construction of structural information Injecting structural information into the aforementioned multiple-choice QA model serves as another auxiliary task. Its input is the question, and its output is structural information. Events are sorted according to their position, and event coreference and causal relationships are combined in a natural language manner. Using the causal graph structure linearization method, the event causal graph structure described in natural language is obtained. (iv) Model training and optimization The arguments generated by the integrated intelligent model and the structural information constructed are injected into the multiple-choice QA model for multiple... Task training involves fine-tuning the trained multiple-choice QA model. The same generated model learns three tasks simultaneously and is trained to minimize the loss. The final multiple-choice QA model is used to identify the causal relationship of events. The same generated model learns three tasks simultaneously and is trained using equations (b) to (e) below to minimize the following label prediction loss: ； in, Indicates the loss weights between tasks. It is the cross-entropy loss function between the predicted token and the target token. This represents a concatenation of custom prompts and questions, used as input to the model; Label the correct answer. Arguments generated for intelligent models For the structural information being constructed; subscript Represents a specific task. For task sets; Corresponding to losses hint Used for output respectively .

Images